Pith. sign in

REVIEW 4 major objections 6 minor 44 references

Mars Traversability Prediction: A Multi-modal Self-supervised Approach for Costmap Generation

T0 review · 4 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A self-supervised IMU-labeling pipeline can train a LiDAR-plus-camera costmap predictor for Mars terrain that keeps error low even when sensors are corrupted, because geometry, not image semantics, drives the learned cost.

desk verdict Honest but flawed: the paper's candid negative result on image semantics is worth something, but the IMU label formula is physically inconsistent and the low MAE claims are uninterpretable until it is fixed. read the letter →

arxiv 2509.11082 v1 pith:3B4QFB5P submitted 2025-09-14 cs.CV cs.RO

Zongwu Xie , Kaijie Yun , Yang Liu , Yiming Ji , Han Li This is my paper
classification cs.CVcs.RO
keywords traversabilitycostmapself-supervisedlearningIMUlabelingmulti-modalfusionplanetaryroverLiDARBEVDINOv3featuressimulation-to-real
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that traversability costmaps for planetary rovers can be learned without human labels by using IMU vibrations and rotations as a self-supervised ground truth, and that a multi-modal network fusing LiDAR geometry with camera features can predict those costmaps reliably. The authors show through ablations that removing image color or the entire image branch barely changes prediction error (MAE stays around 0.08 on a 0–1 scale), while sparsifying LiDAR or adding noise raises error only modestly. From this they conclude that the IMU-derived labels mostly encode terrain geometry, so the LiDAR branch carries the signal and the image branch acts as a minor conditioning input. The central contribution is therefore a reproducible training recipe—simulation from HiRISE terrain models plus IMU-labeled costmaps—rather than a large accuracy gain from multimodality.

What carries the argument

The central mechanism is the IMU-based traversability cost label: TC = w1*sqrt(mean(amag^2)) + w2*theta_cum + w3*sqrt(mean(jmag^2)), where amag is the 3D acceleration magnitude, theta_cum is the cumulative angular change per distance, and jmag is the spatial-domain jerk, with all weights set to 1. This converts raw inertial measurements into a continuous per-cell cost that serves as both training supervision and evaluation metric. Around this, the model uses PointPillars to convert the LiDAR point cloud into a BEV pseudo-image, a ResNet-FPN backbone to extract multi-scale features, a DINOv3 image encoder whose embedding is projected into FiLM modulation parameters to condition the BEV featur

What would settle it

Collect real rover data on two terrains with identical roughness statistics but different slide or sinkage risk (e.g., dry sand versus compacted soil with the same bump profile) and compare the IMU-derived TC labels against an independent hazard measure such as wheel slip ratio or consumed energy; if the labels fail to separate these terrains while the model achieves low MAE against them, the central claim about reliable traversability prediction is not supported.

Watch

Extended reading notes

Core claim

The paper claims that a costmap prediction network trained with IMU-derived labels can produce dense bird's-eye-view traversability maps with mean absolute error around 0.0775 on a normalized scale, and that this error stays below 0.092 even when the LiDAR is randomly thinned by 30% or Gaussian noise is added to both sensors. The driving evidence is that ablations removing image color or the entire learned image encoder leave performance essentially unchanged, which the authors interpret as showing the IMU labels reflect physical bumps and slippage rather than semantic appearance. The paper also claims that the proposed pipeline—a Mars terrain simulator built from HiRISE data, a self-supervi

Load-bearing premise

The load-bearing premise is that the hand-set IMU traversability cost formula, with all three weights fixed at 1, is a faithful measure of true terrain difficulty; if this quantity does not correspond to actual rover hazard, then the model's low prediction error is measuring the wrong thing.

Editorial extensions

If this is right

  • If the central claim holds, planetary rover navigation can use costmaps trained entirely from simulation and IMU feedback, avoiding the need for manual terrain labeling on Mars.
  • The observed robustness to input corruption suggests the model can continue operating acceptably under partial sensor failure, which is directly relevant to long-duration rover missions.
  • The finding that image semantics add little signal implies that for roughness-dominated terrains, a LiDAR-only system with a simple IMU labeling pipeline may be sufficient, reducing computational load onboard.
  • The public simulation environment and labeling pipeline, if released, would let other groups generate comparable Mars-terrain datasets without field access, accelerating research on costmap learning.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension is to collect IMU labels on terrains that are visually different but mechanically similar (e.g., packed sand versus fine gravel) and check whether the predicted costmaps diverge; if they do not, the label formula is missing a hazard component such as slip propensity that does not show up as vibration.
  • The fixed equal weights in the TC formula are a degree of freedom the paper does not explore; learning these weights from a downstream metric (e.g., planning success or wheel slip) could make the supervision more faithful to actual rover risk without changing the architecture.
  • The paper's robustness result is demonstrated on one terrain distribution; a natural stress test is to apply the model to a real Mars-analog field site with substantially different rock density or soil type and measure whether the MAE stays in the 0.08–0.09 range or degrades, which would indicate how much of the apparent robustness is due to distribution similarity.
  • If image semantics are truly irrelevant for IMU-based labels, then a simpler single-modal architecture using only LiDAR geometry could likely match the full model; the paper's ablations point in that direction but stop short of testing a pure-geometry variant at equal capacity.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes a self-supervised framework for predicting traversability costmaps for Mars rovers. A high-fidelity HiRISE-based Gazebo simulation provides RGB, LiDAR, and IMU data; IMU measurements are converted into continuous traversability costs using a hand-defined formula; a PointPillars/ResNet-FPN LiDAR BEV backbone is fused with DINOv3 image features via FiLM and trained with Huber plus smoothness losses. The evaluation is an ablation study on fixed real-world routes, reporting MAE/MSE values around 0.08 on a [0,1] scale, with the main finding that geometry dominates and image semantics provide little benefit. The paper honestly discusses limitations, but the central quantitative claims rest on the validity of the IMU-derived cost labels and on a comparison protocol with no external baselines.

Significance. If the IMU labeling pipeline and costmap predictions are taken at face value, the paper offers a reproducible simulation environment, a scalable self-supervised labeling idea, and a cleanly ablated multi-modal architecture. The authors are commendably candid that the image branch adds little, and the robustness ablations (occlusion, sparsification, noise) are a useful stress test. However, the scientific significance is conditional on two unresolved issues: the IMU-derived ground-truth cost must be physically meaningful and externally validated, and the reported MAE values must be interpretable against a baseline. As presented, the paper's main claim of 'reliable traversability maps' is not yet supported because the supervision signal and evaluation metric may be measuring an arbitrary rescaling of a dimensionally inconsistent heuristic.

major comments (4)
  1. [Section III-C, TC definition] The traversability cost TC = w1*sqrt(mean(amag^2)) + w2*theta_cum + w3*sqrt(mean(jmag^2)) with w1=w2=w3=1 is the sole supervision signal and the reference for all MAE/MSE values. As written, it sums quantities with incompatible units: amag is m/s^2 (and if gravity compensation is not applied, it is dominated by the ~9.8 m/s^2 gravity component, which is nearly constant over terrain), theta_cum is rad/m, and jmag is 1/s^2. The paper never states whether gravity is compensated, how the terms are normalized before summation, or how the raw TC is transformed into the [0,1] labels used in training. Without this, the reported MAE of 0.0775 has no well-defined scale and the central claim of low prediction error is uninterpretable. This must be fixed by specifying preprocessing, per-term normalization or calibration, and by validating the labels against an external measure such as wheel slip, pl
  2. [Section IV-C, Table I] All conclusions about robustness are drawn from single MAE/MSE numbers with no variance estimates or repeated runs. The claim that 'MAE increases from ~0.0775 to 0.0915' when LiDAR is sparsified is a difference of 0.014 that could easily be within run-to-run noise. Moreover, there is no comparison to prior self-supervised costmap methods (e.g., [5], [7], [28]) or even to a simple geometric baseline such as local slope/roughness from the LiDAR point cloud. Without baselines and error bars, the absolute and relative robustness claims are not established. I request mean/std over at least 3-5 seeds and at least one classical or prior-learning baseline.
  3. [Section IV-A and IV-C] The training and evaluation protocol is ambiguous and potentially circular. The method section describes training in a simulated Mars environment (Sec. III-B), while Sec. IV-C describes evaluation on 'fixed test routes in real-world environments' using the physical rover shown in Fig. 1. The dataset section (Sec. IV-A) only mentions the HiRISE dataset and gives no details of the real-world test set, its size, terrain properties, or how labels were generated there. In addition, both training and evaluation use the same hand-defined IMU TC formula, so the reported MAE measures the model's consistency with an internal heuristic, not agreement with any external notion of traversability. The paper should state clearly which data are used for training and testing, and should include at least one external validation (e.g., measured wheel slip, mission-style planning cost, or a held-out human-an
  4. [Section III-C, jerk definition] The quantity j_i = Δa_i / Δs_i is called 'spatial-domain jerk', but its units are s^{-2}, not m/s^3 as conventional jerk. This is not merely a naming issue: the formula suggests a physical meaning that the quantity does not have, and it reinforces the dimensional inconsistency of the TC sum. Please rename or redefine the term, or justify it as a heuristic roughness measure with clearly stated units and normalization.
minor comments (6)
  1. [Abstract/Introduction] The abstract and introduction claim a 'high-fidelity, reproducible simulation environment' and a public dataset, but no dataset URL or release mechanism is provided. If the contribution is reproducibility, the artifacts should be available or a clear commitment to release them should be stated.
  2. [Section III-D, Input Representation] The text mentions 'FAST-LIVO2' here but 'FAST-LIO2' in the Introduction; please ensure consistent naming and correct references.
  3. [Section IV-A] 'The public dataset were utilized' is ungrammatical, and the section does not describe how the HiRISE data were split or processed into the 1,432 trajectory segments mentioned in Sec. III-C.
  4. [Table I] The table would be easier to interpret with a column indicating whether each ablation is a training-time or test-time modification, and with confidence intervals or standard deviations. Also, 'No-Color-PointCloud (train)' surprisingly improves MAE; this should be discussed in the text, not just attributed to overfitting.
  5. [References] Reference [24] has a malformed URL ('httpsarxiv.orgabs2305.08510' instead of 'https://arxiv.org/abs/2305.08510'). Please check all references for consistent formatting.
  6. [General] There are several typos, e.g., 'supported by by the' in the footnote. A careful proofread would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the IMU-derived TC is an explicit proxy label; the model is learned from independent sensors and evaluated on held-out routes.

full rationale

The paper's derivation chain is: IMU measurements are converted to a traversability cost TC by the explicit formula in Sec. III-C (with fixed weights w1=w2=w3=1); those TC values are used as regression targets for the network f(P,I); and the same TC definition is used to compute MAE/MSE on fixed test routes in Sec. IV-C. This is a standard self-supervised proxy-learning setup, not a circular derivation. The predicted costmap is not equal to the TC formula by construction: it is a learned function of LiDAR and RGB inputs, evaluated on held-out routes, and the ablation experiments genuinely perturb the inputs. The weights in the TC formula are fixed rather than fitted, so no fitted parameter is being renamed as a prediction. No load-bearing argument depends on the authors' own prior citations; DINOv3, PointPillars, FiLM, and FAST-LIO2 are external, checkable components. The absence of an external benchmark (e.g., wheel slip, planning outcomes, or human labels) is a validity/grounding limitation of the IMU heuristic, and the paper itself acknowledges this constraint in Sec. V ('The learned model is currently constrained by the IMU-label characteristics and a relatively narrow training set'); however, this is not circularity under the definitions here, because the model output is not defined in terms of the label, nor is the label derived from the model. The dimensional and normalization concerns raised by the skeptic affect whether TC is a meaningful ground truth, but they are correctness/validity issues rather than evidence that the derivation reduces to its own inputs.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the assumption that the IMU-derived TC formula is a valid measure of traversability, and that simulated rigid-body terrain captures real wheel-terrain interaction. These are domain assumptions, not standard math. The free parameters are the hand-set weights and kernel radius. No new physical entities are proposed.

free parameters (3)
  • TC weights w1, w2, w3 = 1.0 each
    Hand-set equal weights in the traversability cost formula (Section III-C). No fitting or sensitivity analysis is reported, yet the entire label and evaluation depend on this choice.
  • Kernel radius r = 1.0 m
    Radius for Bayesian kernel interpolation of TC labels to 0.01m grid (Section III-C). Arbitrary choice affecting label spatial smoothness.
  • Smoothness loss weight lambda_s = 0.1
    Weight balancing the smoothness regularizer in the total loss (Section III-F). Chosen by hand.
assumptions (3)
  • domain assumption IMU-derived acceleration, jerk, and angular change are valid proxies for terrain traversability.
    The entire supervised signal and evaluation metric are built on this heuristic TC formula (Section III-C). No independent validation is provided.
  • domain assumption Rigid-body terrain models from HiRISE DTMs in Gazebo adequately simulate wheel-terrain contact.
    Section III-B states all terrain elements are rigid bodies, simplifying surface interaction. This may be insufficient for deformable soil, which the paper itself lists as a limitation.
  • domain assumption FAST-LIVO2 provides accurate spatiotemporal alignment of camera, LiDAR, and IMU.
    Mentioned in Section III-A as ensuring alignment. The accuracy of this alignment is taken as given and is load-bearing for the fusion model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mars Traversability Prediction: A Multi-modal Self-supervised Approach for Costmap Generation." pith.science (2026). https://pith.science/paper/3B4QFB5P

@misc{pith2026250911082,
  author       = {Pith},
  title        = {Pith review of: Mars Traversability Prediction: A Multi-modal Self-supervised Approach for Costmap Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3B4QFB5P}},
  note         = {Machine review of arXiv:2509.11082}
}
read the original abstract

We present a robust multi-modal framework for predicting traversability costmaps for planetary rovers. Our model fuses camera and LiDAR data to produce a bird's-eye-view (BEV) terrain costmap, trained self-supervised using IMU-derived labels. Key updates include a DINOv3-based image encoder, FiLM-based sensor fusion, and an optimization loss combining Huber and smoothness terms. Experimental ablations (removing image color, occluding inputs, adding noise) show only minor changes in MAE/MSE (e.g. MAE increases from ~0.0775 to 0.0915 when LiDAR is sparsified), indicating that geometry dominates the learned cost and the model is highly robust. We attribute the small performance differences to the IMU labeling primarily reflecting terrain geometry rather than semantics and to limited data diversity. Unlike prior work claiming large gains, we emphasize our contributions: (1) a high-fidelity, reproducible simulation environment; (2) a self-supervised IMU-based labeling pipeline; and (3) a strong multi-modal BEV costmap prediction model. We discuss limitations and future work such as domain generalization and dataset expansion.

Figures

Figures reproduced from arXiv: 2509.11082 by the authors.

Figure 1
Figure 1. Real-world data collection and traversability prediction. The proposed system is validated on a physical four-wheel rover platform operat￾ing in outdoor unstructured environments. Top: the rover navigating forest terrain. Bottom: multi-modal perception results, including RGB imagery, LiDAR point cloud with estimated trajectory, and the predicted traversability costmap. This setup demonstrates the feasibility of gene… view at source ↗
Figure 2
Figure 2. High-fidelity Mars simulation environment. A photorealistic simulation environment is constructed using HiRISE digital terrain models and texture maps. The simulated rover is equipped with virtual sensors (LiDAR, camera, IMU), allowing scalable data collection and controllable evaluation across diverse Martian-like terrains. Insets show the elevation profile and simulated LiDAR scan, providing a realistic approximat… view at source ↗
Figure 3
Figure 3. Overview of the proposed traversability learning framework. The pipeline consists of three main modules: (a) rover and environment, where a four-wheel rover equipped with LiDAR, RGB camera, and IMU is deployed in Mars-analog terrains; (b) SLAM and TC label generation, where multi-modal sensor data are temporally aligned via FAST-LIO2 and IMU-based metrics are used to derive traversability cost ground truth; and (c) … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 1 canonical work pages

  1. [5]

    How does it feel? self-supervised costmap learning for off-road vehicle traversability,

    M. G. Castro, S. Triest, W. Wang, J. M. Gregory, F. Sanchez, J. G. Rogers, and S. Scherer, “How does it feel? self-supervised costmap learning for off-road vehicle traversability,” in2023 IEEE International Conference on Robotics and Automation (ICRA), 2023, pp. 931–938

    2023

  2. [7]

    Learning risk-aware costmaps for traversability in challenging environments,

    D. D. Fan, A.-A. Agha-Mohammadi, and E. A. Theodorou, “Learning risk-aware costmaps for traversability in challenging environments,” IEEE robotics and automation letters, vol. 7, no. 1, pp. 279–286, 2021

    2021

  3. [28]

    Badgr: An autonomous self- supervised learning-based navigation system,

    G. Kahn, P. Abbeel, and S. Levine, “Badgr: An autonomous self- supervised learning-based navigation system,”IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 1312–1319, 2021

    2021

  4. [1]

    Traversability analysis and path planning for a planetary rover,

    D. B. Gennery, “Traversability analysis and path planning for a planetary rover,”Autonomous Robots, vol. 6, no. 2, pp. 131–146, 1999

    1999

  5. [2]

    Stereo vision and rover navigation software for planetary exploration,

    S. Goldberg, M. Maimone, and L. Matthies, “Stereo vision and rover navigation software for planetary exploration,” inProceedings, IEEE Aerospace Conference, vol. 5, 2002, pp. 5–5

    2002

  6. [3]

    Obstacle de- tection and terrain classification for autonomous off-road navigation,

    R. Manduchi, A. Castano, A. Talukder, and L. Matthies, “Obstacle de- tection and terrain classification for autonomous off-road navigation,” Autonomous robots, vol. 18, no. 1, pp. 81–102, 2005

    2005

  7. [4]

    Traversability index: a new concept for planetary rovers,

    H. Seraji, “Traversability index: a new concept for planetary rovers,” inProceedings 1999 IEEE International Conference on Robotics and Automation (Cat. No.99CH36288C), vol. 3, 1999, pp. 2006–2013 vol.3

    1999

  8. [6]

    Tartandrive: A large-scale dataset for learning off-road dynamics models,

    S. Triest, M. Sivaprakasam, S. J. Wang, W. Wang, A. M. Johnson, and S. Scherer, “Tartandrive: A large-scale dataset for learning off-road dynamics models,” in2022 International Conference on Robotics and Automation (ICRA). IEEE, 2022, pp. 2546–2552

    2022

Show all 44 references
  1. [8]

    Learning-based traversability costmap for autonomous off- road navigation,

    Q. Zhu, Z. Sun, S. Xia, G. Liu, K. Ma, L. Pei, Z. Gong, and C. Jin, “Learning-based traversability costmap for autonomous off- road navigation,” inChina Intelligent Robotics Annual Conference. Springer, 2024, pp. 301–312

    2024

  2. [9]

    A reliable traversability learning method based on human-demonstrated risk cost mapping for mobile robots over uneven terrain,

    B. Zhang, G. Li, J. Zhang, and X. Bai, “A reliable traversability learning method based on human-demonstrated risk cost mapping for mobile robots over uneven terrain,”Engineering Applications of Artificial Intelligence, vol. 138, p. 109339, 2024

    2024

  3. [10]

    A survey on terrain traversability analysis for autonomous ground vehicles: Methods, sensors, and challenges,

    P. V . K. Borges, T. Peynot, S. Liang, B. Arain, M. Wildie, M. G. Minareci, S. Lichman, G. Samvedi, I. Sa, N. Hudson, M. Milford, P. Moghadam, and P. Corke, “A survey on terrain traversability analysis for autonomous ground vehicles: Methods, sensors, and challenges,” Field Ro...

    2022

  4. [11]

    Vision-based terrain characterization and traversability assessment,

    A. Howard and H. Seraji, “Vision-based terrain characterization and traversability assessment,”Journal of Robotic Systems, vol. 18, no. 10, pp. 577–587, 2001. [Online]. Available: https://onlinelibrary.wiley.com/doi/abs/10.1002/rob.1046

    2001 doi

  5. [12]

    Texture-based segmentation for sand and rock in mars images,

    O. Alkawi, R. Qahwaji, C. Smith, and Z. Mnasri, “Texture-based segmentation for sand and rock in mars images,”Advances in Space Research, vol. 76, no. 3, pp. 1756–1773, 2025. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0273117725004934

    2025

  6. [13]

    Real- time semantic mapping for autonomous off-road navigation,

    D. Maturana, P.-W. Chou, M. Uenoyama, and S. Scherer, “Real- time semantic mapping for autonomous off-road navigation,” inField and Service Robotics: Results of the 11th International Conference. Springer, 2017, pp. 335–350

    2017

  7. [14]

    Review of autonomous path planning for planetary rovers,

    Z. Haojie, J. Feng, L. Chuankai, Z. Zuoyu, and L. Qing, “Review of autonomous path planning for planetary rovers,”Chinese Journal of Engineering, vol. 46, no. 11, pp. 2063–2075, 2024. [Online]. Available: http://cje.ustb.edu.cn/en/article/doi/10.13374/j.issn2095- 9389.2024.01.04.001

  8. [15]

    Challenges and solutions for autonomous ground robot scene understanding and navigation in unstructured outdoor environments: A review,

    L. Wijayathunga, A. Rassau, and D. Chai, “Challenges and solutions for autonomous ground robot scene understanding and navigation in unstructured outdoor environments: A review,” Applied Sciences, vol. 13, no. 17, 2023. [Online]. Available: https://www.mdpi.com/2076-3417/13/17/9877

    2023

  9. [16]

    A photorealistic terrain simulation pipeline for unstructured outdoor environments,

    M. G. M ¨uller, M. Durner, A. Gawel, W. St¨urzl, R. Triebel, and R. Sieg- wart, “A photorealistic terrain simulation pipeline for unstructured outdoor environments,” in2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021, pp. 9765–9772

    2021

  10. [17]

    Rca: Ride comfort-aware visual navi- gation via self-supervised learning,

    X. Yao, J. Zhang, and J. Oh, “Rca: Ride comfort-aware visual navi- gation via self-supervised learning,” in2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022, pp. 7847– 7852

    2022

  11. [18]

    V-strong: Visual self-supervised traversability learning for off-road navigation,

    S. Jung, J. Lee, X. Meng, B. Boots, and A. Lambert, “V-strong: Visual self-supervised traversability learning for off-road navigation,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 1766–1773

    2024

  12. [19]

    Follow the footprints: Self- supervised traversability estimation for off-road vehicle navigation based on geometric and visual cues,

    Y . Jeon, E. I. Son, and S.-W. Seo, “Follow the footprints: Self- supervised traversability estimation for off-road vehicle navigation based on geometric and visual cues,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 1774– 1780

    2024

  13. [20]

    Roadrunner—learning traversability estimation for autonomous off-road driving,

    J. Frey, M. Patel, D. Atha, J. Nubert, D. Fan, A. Agha, C. Pad- gett, P. Spieler, M. Hutter, and S. Khattak, “Roadrunner—learning traversability estimation for autonomous off-road driving,”IEEE Transactions on Field Robotics, vol. 1, pp. 192–212, 2024

    2024

  14. [21]

    Lidar- inertial slam with dem-driven global constraints for planetary rover exploration,

    X. Zhang, Y . Li, Z. Cao, J. Lv, Z. Huang, and W. Zhang, “Lidar- inertial slam with dem-driven global constraints for planetary rover exploration,”The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, vol. XLVIII- 3-2024, pp. 615–62...

    2024

  15. [22]

    Fully convolutional networks for semantic segmentation,

    J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks for semantic segmentation,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2015, pp. 3431–3440

    2015

  16. [23]

    Learning deconvolution network for semantic segmentation,

    H. Noh, S. Hong, and B. Han, “Learning deconvolution network for semantic segmentation,” inProceedings of the IEEE international conference on computer vision, 2015, pp. 1520–1528

    2015

  17. [24]

    Fast traversability estimation for wild visual navigation,

    J. Frey, M. Mattamala, N. Chebrolu, C. Cadena, M. Fallon, and M. Hutter, “Fast traversability estimation for wild visual navigation,”

  18. [25]

    Wayfast: Navigation with predictive traversability in the field,

    M. V . Gasparino, A. N. Sivakumar, Y . Liu, A. E. B. Velasquez, V . A. H. Higuti, J. Rogers, H. Tran, and G. Chowdhary, “Wayfast: Navigation with predictive traversability in the field,”IEEE Robotics and Automation Letters, vol. 7, no. 4, pp. 10 651–10 658, 2022

    2022

  19. [26]

    Active traversability learning via risk- aware information gathering for planetary exploration rovers,

    M. Endo and G. Ishigami, “Active traversability learning via risk- aware information gathering for planetary exploration rovers,”IEEE Robotics and Automation Letters, vol. 7, no. 4, pp. 11 855–11 862, 2022

    2022

  20. [27]

    Mlnav: Learning to safely navigate on martian terrains,

    S. Daftry, N. Abcouwer, T. D. Sesto, S. Venkatraman, J. Song, L. Igel, A. Byon, U. Rosolia, Y . Yue, and M. Ono, “Mlnav: Learning to safely navigate on martian terrains,”IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 5461–5468, 2022

    2022

  21. [29]

    Learning to drive off road on smooth terrain in unstructured environments using an on-board camera and sparse aerial images,

    T. Manderson, S. Wapnick, D. Meger, and G. Dudek, “Learning to drive off road on smooth terrain in unstructured environments using an on-board camera and sparse aerial images,” in2020 IEEE International Conference on Robotics and Automation (ICRA), 2020, pp. 1263– 1269

    2020

  22. [30]

    Risk-aware off-road navigation via a learned speed distribution map,

    X. Cai, M. Everett, J. Fink, and J. P. How, “Risk-aware off-road navigation via a learned speed distribution map,” in2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022, pp. 2931–2937

    2022

  23. [31]

    Sim ´eoni, H

    O. Sim ´eoni, H. V . V o, M. Seitzer, F. Baldassarre, M. Oquab, C. Jose, V . Khalidov, M. Szafraniec, S. Yi, M. Ramamonjisoa, F. Massa, D. Haziza, L. Wehrstedt, J. Wang, T. Darcet, T. Moutakanni, L. Sentana, C. Roberts, A. Vedaldi, J. Tolan, J. Brandt, C. Couprie, J. Mairal, H...

    2025 arXiv

  24. [32]

    Learning off-road terrain traversability with self-supervisions only,

    J. Seo, S. Sim, and I. Shim, “Learning off-road terrain traversability with self-supervisions only,”IEEE Robotics and Automation Letters, vol. 8, no. 8, pp. 4617–4624, 2023

    2023

  25. [33]

    Semantic terrain classification for off-road autonomous driving,

    A. Shaban, X. Meng, J. Lee, B. Boots, and D. Fox, “Semantic terrain classification for off-road autonomous driving,” inConference on Robot Learning, 2021. [Online]. Available: https://api.semanticscholar.org/CorpusID:244037243

    2021

  26. [34]

    Ga-nav: Efficient terrain segmenta- tion for robot navigation in unstructured outdoor environments,

    T. Guan, D. Kothandaraman, R. Chandra, A. J. Sathyamoorthy, K. Weerakoon, and D. Manocha, “Ga-nav: Efficient terrain segmenta- tion for robot navigation in unstructured outdoor environments,”IEEE Robotics and Automation Letters, vol. 7, no. 3, pp. 8138–8145, 2022

    2022

  27. [35]

    Terrain segmentation and roughness estimation using rgb data: Path planning application on the centauro robot,

    V . Suryamurthy, V . S. Raghavan, A. Laurenzi, N. G. Tsagarakis, and D. Kanoulas, “Terrain segmentation and roughness estimation using rgb data: Path planning application on the centauro robot,” in 2019 IEEE-RAS 19th International Conference on Humanoid Robots (Humanoids), 201...

    2019

  28. [36]

    Hybrid terrain traversability analysis in off-road environments,

    T. H. Y . Leung, D. Ignatyev, and A. Zolotas, “Hybrid terrain traversability analysis in off-road environments,” in2022 8th Interna- tional Conference on Automation, Robotics and Applications (ICARA), 2022, pp. 50–56

    2022

  29. [37]

    A mars local terrain matching method based on 3d point clouds,

    B. Wang, S. Zhao, X. Guo, and G. Yu, “A mars local terrain matching method based on 3d point clouds,”Remote Sensing, vol. 16, no. 9, 2024. [Online]. Available: https://www.mdpi.com/2072- 4292/16/9/1620

    2024

  30. [38]

    Fast-livo2: Fast, direct lidar-inertial-visual odometry,

    C. Zheng, W. Xu, Z. Zou, T. Hua, C. Yuan, D. He, B. Zhou, Z. Liu, J. Lin, F. Zhu, Y . Ren, R. Wang, F. Meng, and F. Zhang, “Fast-livo2: Fast, direct lidar-inertial-visual odometry,” 2024. [Online]. Available: https://arxiv.org/abs/2408.14035

    2024 arXiv

  31. [39]

    Bayesian generalized kernel inference for terrain traversability mapping,

    T. Shan, J. Wang, B. Englot, and K. Doherty, “Bayesian generalized kernel inference for terrain traversability mapping,” inConference on robot learning. PMLR, 2018, pp. 829–838

    2018

  32. [40]

    Pointpillars: Fast encoders for object detection from point clouds,

    A. H. Lang, S. V ora, H. Caesar, L. Zhou, J. Yang, and O. Beijbom, “Pointpillars: Fast encoders for object detection from point clouds,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 12 697–12 705

    2019

  33. [41]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

    2016

  34. [42]

    Feature pyramid networks for object detection,

    T.-Y . Lin, P. Doll´ar, R. Girshick, K. He, B. Hariharan, and S. Belongie, “Feature pyramid networks for object detection,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 2117–2125

    2017

  35. [43]

    Film: Visual reasoning with a general conditioning layer,

    E. Perez, F. Strub, H. De Vries, V . Dumoulin, and A. Courville, “Film: Visual reasoning with a general conditioning layer,” inProceedings of the AAAI conference on artificial intelligence, vol. 32, no. 1, 2018

    2018

  36. [2023]

    Available: httpsarxiv.orgabs2305.08510

    [Online]. Available: httpsarxiv.orgabs2305.08510

    arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.